Rising component costs may adjust prices on selected items without prior notice. Supply constraints may add shipping time — we’ll notify you if affected. OEM costs and supply tightening — contact us to secure stock and pricing. For projects or bulk orders, choose “Request for Quotation” on any item.
Call our specialists(65) 6978 3502 Email our sales teamsales@sourceit.com.sg

Networking & Cybersecurity

Copper, Optical or Co-Packaged? The Cables That Decide How Far an AI Cluster Can Stretch

Nobody puts cables on the cover of an AI brochure. Yet the cable decides how far apart two machines can sit, how much power the network burns, and how often someone is on a ladder replacing a failed part.

There are five options, and the industry is in the middle of an argument about the fifth. The short version: use the simplest cable that reaches. The rest of this article explains why.

The five options at a glance

Type What it is Typical reach Power
🟤 DAC
Direct attach copper
A plain copper cable with a plug on each end Up to about 3 m
Almost none
🟠 AEC
Active electrical cable
Copper with small chips in the plugs that clean up the signal About 3 to 7 m
Low
🔵 AOC
Active optical cable
Fibre with the optics sealed into the plugs Up to about 100 m
Medium
🟣 Pluggable optics
Transceiver plus fibre
A separate optical module at each end, joined by ordinary fibre 100 m to kilometres
Highest
⚪ CPO
Co-packaged optics
The optics are built into the switch chip package; fibre plugs straight into the switch Optical distances Lower than pluggables
Vendor claim

Sources: Network-Switch: DAC vs AOC vs optical transceivers, Axiom: DAC vs AOC vs AEC vs ACC. Reach figures are typical values and shrink as speeds rise; check the specification for your speed. The bars are indicative, not to scale.

Why copper keeps winning inside the rack

A passive copper cable has no electronics. It uses almost no power, adds no processing delay and has very little that can fail. That is why the densest AI racks use copper for the short links between GPUs and switches wherever the distance allows.

The catch is physics. The faster the signal, the shorter the distance copper can carry it. At today's speeds, passive copper is good for a few metres at most, which is roughly one rack.

AEC: the quiet middle option

Active electrical cables add small chips that recondition the signal, stretching copper to roughly 3 to 7 metres. That is enough to reach the next rack or two without paying for optics. One supplier's comparison puts AEC power at 25 to 50% below an equivalent optical cable. For a small cluster of adjacent racks, AEC is often the option nobody mentioned and the one that fits best.

Optics: when distance leaves no choice

Beyond a few metres, light takes over. An active optical cable is the simple form: fibre with the optics sealed in, plug and play, good for runs of tens of metres. Pluggable transceivers with separate fibre go much further and let you reuse the fibre when speeds change.

The price is power and parts. Pluggable modules are commonly quoted at a few watts each. That sounds small until it is multiplied across thousands of ports, and every module is a component that can fail.

Co-packaged optics: the argument of the moment

Co-packaged optics removes the pluggable module by building the optical parts into the switch itself. NVIDIA announced CPO switches in two families: Quantum-X Photonics for InfiniBand and Spectrum-X Photonics for Ethernet. Its headline claims are 3.5 times better power efficiency and 10 times better resiliency than traditional designs. NVIDIA's developer blog gave commercial availability as early 2026 for the InfiniBand switches and the second half of 2026 for Ethernet.

The case for: at very large scale, network optics become a serious share of the power bill, and fewer discrete parts should mean fewer failures.

The case for caution: the numbers are NVIDIA's own. Analysts quoted when the switches were announced pointed to open questions about the technology's maturity and its business model. There is also a practical one: with a pluggable module, a failed optic is swapped in a minute. When the optics are part of the switch, the repair is a different conversation.

Sources: NVIDIA announcement, NVIDIA developer blog on co-packaged optics, Fierce Network

Which cable for which link

The link Start with
🗄️ Server to switch in the same rack DAC
↔️ To the next rack or two AEC
🏢 Across a row or a room AOC or pluggable optics
🏙️ Between halls or buildings Pluggable optics
🏭 A new build with thousands of GPUs Ask vendors about CPO, and ask who repairs it

Three practical points

  • Measure the real route. A cable runs up, across and down, not in a straight line. A two-metre gap between ports can need a three-metre cable.
  • Check compatibility. Switch and network card vendors publish lists of supported cables. An unlisted cable can work on the bench and misbehave under load.
  • Order spares. Cables are cheap next to the GPUs waiting on them.

For most clusters under a few racks, the honest answer is undramatic: copper where it reaches, optics where it does not. The interesting arguments belong to people building at a scale most buyers never will.

Related on SourceIT

Share This Article
Facebook

More from our blog


← Older Post Newer Post →